Intern(Chow Jun Hoe)

 31/7/2025 Today, my focus is to verify the impact of dummying future data's time-category variables on the accuracy of the XGBoost model. After multiple rounds of testing, the results show that this adjustment indeed improves model performance during certain time periods. In particular, the discrepancy between the predicted curve and actual sales data is significantly reduced, indicating that the model is better able to capture sales trends after price changes.

 30/7/2025 This morning, I attended the Financial Year 2025 presentation meeting, which detailed Eco-Shop's achievements over the past year and outlined key goals for 2026. In the afternoon, I returned to my XGBoost model research, experimenting with future data. This time, I tested whether adding get_dummies (one-hot encoding of categorical variables) to future data could further improve prediction accuracy. Specifically, encoding holiday, month, and day of the week information for future dates would allow the model to better identify future time characteristics.

 29/7/2025 This morning, I had planned to continue testing and optimizing the XGBoost model, specifically adjusting parameters to account for holiday factors and lag features. However, my work schedule suddenly changed when my company informed me that I needed to assist with the handling of supplies for a charity event sponsored by Eco-Shop. As a result, I spent almost the entire day supporting this effort.

 28/7/2025 Today, my work focuses on re-adjusting the time range around festivals to further optimize the model's performance in Eco-Shop sales forecasting, with the goal of reducing MAPE and making the forecast results closer to actual sales trends. In the previous version, the model was already able to capture the sales on Hari Raya, but after observing the results, it was found that Awal Ramadan would also cause an increase in sales, so I tried to add this factor in.

 25/7/2025 Today's model adjustment work has achieved interim results. Through carefully designed holiday-related variables, I successfully made the XGBoost model's prediction trend during Hari Raya closer to the actual sales data. Not only did it show a significant expected increase before the festival, but it also correctly captured the peak during the festival and the decline after the festival. The overall accuracy (such as MAPE) has also improved. However, when further observing the model's prediction performance at other key points (such as after the product price increase), it was found that the model's perception of "sales volume changes after the price increase" is insufficient. Further adjustments are needed.

 24/5/2025 In the previous model building, I found that after adding lag_365, lag_730 and other lag features, the model predicted well in normal time periods, but around holidays, especially during peak consumption periods like Hari Raya, the prediction effect was biased. I strengthened the dummy features related to festivals in the model, such as marking the day of Hari Raya (is_raya), the day before (is_raya_before) and the day after (is_raya_after) as independent variables to help the model clearly distinguish between festival periods.

 23/7/2025 In terms of the forecast model, today I further added the lag feature (lagging sales). The experimental results show that the forecast line is closer to the actual sales, especially for non-holiday sales. However, an important problem was also found: the effect of the lag feature will mask the impact of external factors such as holidays. For example, even if there are promotions on the eve of certain holidays, the model will refer to the previous "non-holiday" sales value, resulting in insufficient prediction of the peak period. At 10:30 this morning, the Prime Minister made an important statement. At the request of the manager, an impact analysis was conducted on the news and presented to him in the afternoon.

 22/7/2025 Today's work focuses on further optimizing the XGBoost model, especially conducting more dimensional exploration and experiments in the feature engineering stage, in order to further reduce the prediction error (MAPE) and improve the generalization ability of the model. I tested and adjusted the expression of periodic changes. The original model only used time fields such as "day" and "month" as categorical features, but this processing method is not sensitive enough to the potential periodicity in the time series (such as the cyclicity of the month or the difference in patterns on different days of the week). Therefore, I began to try to add sine and cosine (sin/cos) conversion to represent periodicity.

 21/7/2025 Today's work mainly focuses on continuing to optimize the XGBoost model, especially making more in-depth adjustments and learning on the handling of holiday factors. In the morning, I reviewed and re-studied the impact of holiday factors on time series forecasting, especially how to transform holiday information into features that can be interpreted by the model in a structured way. In the afternoon, I attended an internal lecture held by the company. The lecture focused on the application trends of artificial intelligence in today's business environment. It also emphasized the attitude changes that employees need to have in the context of rapid technological development, such as continuous learning, cross-departmental collaboration and proactive innovation.

 18/7/2025 The focus of my work this morning was to discuss with my supervisor and team the sales volume forecast results I had previously completed using the XGBoost model. During the discussion, I introduced the features used in the model (including historical sales, holiday markers, sales volume changes before price increases, etc.), model structure, and MAPE performance in detail to my supervisor and team. The supervisor affirmed the overall accuracy of the model, especially compared to the previously used SARIMA model, XGBoost is more stable in handling multiple variables and predicting nonlinear trends. However, the supervisor also made some suggestions for improvement, such as further subdividing the behavioral changes before and after the holidays, and considering the model differences between promotional items and non-promotional items.

 17/7/2025 Today's work focuses on preparing for tomorrow's XGBoost model report and further consolidating the understanding and mastery of the entire prediction script to ensure that the logic, advantages and application effects of the model can be clearly explained in the report. In the afternoon, I participated in the meeting of the Supply Chain department to understand the current status of goods flow and replenishment, as well as the KPI tasks that need to be completed.

 16/7/2025 Today's main work focuses on re-studying and optimizing the Temporal Convolutional Network (TCN) model in order to improve the accuracy of sales volume forecasting by introducing external time series factors. Since the results of the previous TCN model were not ideal, today's task focuses on trying to observe whether the model performance improves by introducing holiday and starting price factors (such as sales volume changes before price adjustment).

 15/7/2025 Today, I rebuilt and optimized the XGBoost model, and further strengthened the business sensitivity of the model based on the original sales forecasting architecture. I successfully added the key factors of "sales change after price increase" and "sales increase one day before the festival", making the model closer to the actual operation scenario of Ecoshop. From the model evaluation results, the overall MAPE has decreased, and the model's explanatory power in actual business scenarios has been enhanced. The prediction curve can better reflect the changing logic of sales behavior.

 14/7/2025 In the morning, I participated in the weekly Promotion Meeting. The meeting focused on the upcoming or ongoing promotional products, the current inventory status, and whether there is a risk of out-of-stock (OOS). In the afternoon, I started to build and experiment with the N-BEATS (Neural Basis Expansion Analysis for Time Series) model as part of exploring a more accurate sales forecasting method.

 11/7/2025 Today I successfully completed the experiment of integrating holiday factors into the TCN model and conducting sales forecasting. The model runs normally in terms of structure and training, and external variables such as holidays are also successfully incorporated into the modeling process. However, when actually evaluating the forecast results, it was found that even after adding holiday variables, the performance of the model still could not adapt well to the sales characteristics of Ecoshop, the forecast error was still high, the trend fluctuated greatly, and it did not meet the conditions for actual deployment. Therefore, after analysis and consideration, it was decided to temporarily end the research and experimental stage of the TCN model and start to explore the next cutting-edge time series forecasting model - N-BEATS (Neural Basis Expansion Analysis for Time Series).

 10/7/2025 Today I officially started to build the TCN (Temporal Convolutional Network) model and tried to apply it to the sales forecasting task. Although I successfully completed the basic construction and training process of the model, when evaluating the forecast results, I found that the model output fluctuated greatly, and there was a significant deviation between the forecast curve and the actual sales, so I was unable to achieve the effect that can be directly applied.

 9/7/2025 Today I started to study and learn TCN (Temporal Convolutional Network). As a new time series prediction model, TCN is regarded as a more stable and parallel-efficient alternative to traditional RNN/LSTM. It is widely used in demand forecasting, inventory management, financial time series and other fields.

 8/7/2025 Today, the focus is on rechecking and understanding the complete script of the XGBoost model, reading and parsing each code line by line, ensuring that not only the model can be used, but also the logic and principles behind it can be clearly understood. After that, continue to expand the exploration of sales forecasting models and start looking for and studying other forecasting methods besides XGBoost and SARIMA in order to find a more stable and accurate forecasting solution.

 7/7/2025 Today, I successfully completed the construction and evaluation of the XGBoost model, and compared its prediction results with the previously established SARIMA model. We finally confirmed that XGBoost is superior in overall performance, especially in terms of prediction accuracy.

 4/7/2025 In the morning, I held a meeting with the Manager and Supply Chain Executive to report and discuss the task assigned by the Manager last time. The task was to conduct an in-depth analysis of 10 designated Item No, covering the inventory status, sales trends, Raw Remark content, and whether there were Reserve or In Transit issues for each product. In the afternoon, i continued to complete the construction of the XGBoost model and further optimized it based on the previous parameter adjustments and the introduction of holiday factors.

 3/7/2025 After completing Auto Parameter Tuning today, i successfully achieved a relatively ideal sales forecast result. On this basis, i further began to introduce holiday effects to enhance the model's ability to capture holiday cyclical fluctuations, making the forecast results closer to the actual business trend.

 2/7/2025 Today, i continue to build and optimize the XGBoost model, focusing on adjusting the rationality of the model performance, such as starting to try to adjust the parameters. The goal is to improve the accuracy of sales forecasts and make the output results more in line with actual business trends.

1/7/2025 Today, I continued to deepen my understanding of the XGBoost model and officially started to build a model for sales forecasting. At the same time, I also completed the new tasks assigned by the Executive to ensure that both tasks progressed steadily.